JP3470028B2 - Handwritten character recognition device and program storage medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwritten character recognition device and a program storage medium used for personal computers, portable information terminals, etc., for recognizing handwritten characters pen-inputted from a tablet.

[0002]

2. Description of the Related Art At present, a method of using a keyboard to input characters to a personal computer, a portable information terminal or the like is the mainstream. However, character input using this keyboard requires a certain amount of skill, and the device for character input is becoming smaller, such as a portable information terminal, so it is possible to secure an area for providing a keyboard. There are problems such as not being. Therefore, online handwritten character input using a tablet or the like has attracted attention.

Generally, writing styles are classified into three types as shown in FIGS. 1 (a) to 1 (c). In FIG. 1, these writing styles are shown in order of complexity of recognition. The first style shown in FIG. 1 (a) allows contact and overlap between characters, but each character is a type of writing separately (hereinafter referred to as writing style). In addition, the second style shown in FIG. 1 (b) is generally "t" after writing letters and letters in a word in succession.
A type that writes down the crossbar, the dot of "i", etc.
(Hereinafter referred to as Cursive style). Also, Fig. 1 (c)
The third style shown in (1) is a type (hereinafter referred to as mixed style) in which the writing style and the cursive style are mixed.

[I] As a system for recognizing the writing style, for example, there is a method of recognizing handwritten characters which may be touched or overlapped by the following procedures (1) to (3) (US Patent No. 4,731,857 (Japanese Patent Laid-Open No. 61-15287). (1) Consider the ends of all strokes in a written word as possible character segmentation positions. (2) All combinations of sections that can be characters (one or more stroke sequences up to the maximum number of strokes) are sent to the character recognition unit to obtain character candidates for the section. (3) The recognized character candidates are combined to generate a word, and the best word is used as the recognition result.

[Ii] Further, as a system for recognizing the Cursive style, there is a method of recognizing handwritten characters continuously written by the following procedures (1) to (3) (US Pat. No. 5,313,527). . (1) Continuously written handwritten characters are expressed using about 30 types of symbols called metastrokes. (2) For each word to be processed, the similarity is calculated by matching the word previously expressed in the metastroke sequence with the input metastroke sequence. (3) The word determined to have the highest similarity is used as the recognition result of the input handwritten character.

[0006]

However, the recognition methods [i] and [ii] described above have the following problems, and only limited character recognition is possible.

First, in the recognition method [i] above, although (a) characters may overlap,
Each character is written separately and ends before the next character begins. (b) Character segmentation derives almost all true segmentation trialable points. Therefore, there is a problem that a writer cannot return to a delayed stroke such as a crossbar of "t" or a dot of "i" written back after writing a word. Moreover, when trying to apply this recognition method to the Cursive style, it is difficult to derive all true segmentation trialable points of the words written in the Cursive style.

In the recognition method [ii], the recognition performance is relatively good when writing in the Cursive style in which the writing method of each character is limited to some extent, but the stroke order variation of each character,
There is a problem that the recognition performance is low for the Mixed style including the deformation of the character shape. The simplest solution to this problem is to increase the variations when expressing each character with a metastroke, and to have a metastroke sequence of words that combine each variation as a dictionary. However, it is difficult to construct a general system because one word dictionary increases by multiplying the variations of each character and the storage capacity increases.

SUMMARY OF THE INVENTION An object of the present invention is to provide a handwritten character recognition device that can recognize words written in the writing style, the cursive style and the mixed style.

[0010]

In order to achieve the above object, the handwritten character recognizing device according to the first aspect of the present invention uses a stroke represented by a coordinate data string in which the coordinates of handwriting input by handwriting are arranged in time series, either one of the sub-stroke, or a stroke segmentation unit for dividing into two or more sub-stroke, upper Symbol sub stroke or the sub stroke obtained by the stroke segmentation unit normalizes each sub stroke chain multiple continuous The normalization unit and the feature point sequence having the feature amount related to the substroke or the feature amount related to the substroke chain normalized by the normalization unit are input into the handwriting input.
The above-mentioned substro obtained from the stroke of the applied handwriting
Over click or a feature pattern of the sub-stroke chain
And a feature extraction unit for extracting and a coordinate data string of a predetermined character or a pattern forming a part of the character
A feature point sequence having, for each sub-pattern, a feature amount associated with a sub-stroke forming the sub-pattern or a feature amount associated with a sub-stroke chain in which a plurality of sub-strokes are connected.
Is obtained from the strokes of the handwriting input by handwriting above.
Registered as a standard pattern for matching with
The sub pattern dictionary storage unit recorded, the dictionary selection unit that selects a standard pattern satisfying a predetermined condition from the standard patterns registered in the sub pattern dictionary storage unit, and the feature extraction unit. By comparing the characteristic pattern with the standard pattern selected from the sub-pattern dictionary storage unit by the dictionary selection unit, from the standard pattern similar to the characteristic pattern,
And a word generation unit for generating a handwritten input word by combining the sub-pattern candidates recognized by the matching unit with reference to a word dictionary. Is characterized by.

According to the handwritten character recognition device of the first aspect, the coordinate data string in which the coordinates of the handwriting input by handwriting are arranged in time series is input from the input device such as the tablet, and the end of one stroke is completed. Each time, the stroke segmentation unit divides the stroke represented by the coordinate data string into one or more substrokes. Then, the coordinate data string of the sub-stroke obtained by the stroke segmentation unit, the sub-stroke that can be a sub-pattern that forms a character or a part of the character Normalize. The sub-pattern may form a character string regardless of one character. The feature extraction unit extracts a feature point sequence having a feature amount related to the substroke or the substroke chain normalized by the normalization unit, as a feature pattern. On the other hand, a standard pattern extracted as a feature point sequence having a feature amount related to a substroke or a substroke of a subpattern that forms a part of a predetermined character or a substroke chain is stored in advance in the subpattern dictionary storage unit. Register In this way, the dictionary selection unit selects a standard pattern satisfying a predetermined condition from the standard patterns registered in the sub-pattern dictionary storage unit. Then, the matching unit collates the characteristic pattern extracted by the characteristic extraction unit with the standard pattern selected from the sub-pattern dictionary storage unit to obtain a sub-pattern candidate. Thus, for each stroke, sub-stroke division, normalization,
Feature extraction, selection of standard patterns, and recognition of sub-pattern candidates are sequentially repeated to recognize sub-pattern candidates that form a character or a part of a character. Then, the word generation unit combines the sub-pattern candidates for each stroke while referring to the word dictionary to generate a handwritten word.

[0012] Thus, the stroke of a handwritten handwriting is made into one substroke or is divided into two or more substrokes, so that the characters existing in the cursive style and mixed style strokes and the continuation of the characters are continued. Most of the parts are separated into individual characters. This character separation is in Cursive style and Mixed
Since it is difficult to separate all the characters in a style stroke, the Cursiv concept is introduced by introducing the concept of a sub-pattern that forms a character or a part of a character to deal with difficult-to-separate character chains and delayed strokes.
Can recognize e-style and Mixed-style handwriting.

The handwritten character recognizing device of claim 2 is
The handwritten character recognition device according to claim 1, wherein the normalization unit selects any one of a plurality of normalization conditions according to an aspect ratio of a coordinate data string of the substroke or the substroke chain. It has a normalizing condition selecting means, and the normalizing section is characterized by performing normalization under the normalizing condition selected by the normalizing condition selecting means.

According to the handwritten character recognition device of the second aspect, the normalization condition selection means of the normalization unit makes a plurality of normalizations according to the aspect ratio of the coordinate data string of the substroke or the substroke chain. The normalization unit selects any one of the conditions, and normalizes the selected normalization condition. Therefore, by performing normalization according to the aspect ratio of the coordinate data string of the substroke or the substroke chain, it is possible to prevent extreme deformation due to normalization of the input pattern.

Further, the handwritten character recognition device according to claim 3 is
The handwritten character recognition device according to claim 2, wherein the sub-pattern dictionary storage unit stores, for each of the registered standard patterns,
It has a header section in which information on the normalization conditions used when creating the standard pattern is registered, and the dictionary selection section stores information on the normalization conditions in the header section of the sub-pattern dictionary storage section. It is characterized in that only the standard patterns that it has are selected from the sub-pattern dictionary storage section.

According to the handwritten character recognition device of the third aspect, information on the normalization condition used when the standard pattern to be registered in the sub-pattern dictionary storage is created is stored in the header of the sub-pattern dictionary storage. It is registered in advance for each standard pattern. Then, by the dictionary selection unit,
From the standard patterns registered in the sub-pattern dictionary storage unit, only the standard pattern having the information on the normalization condition in the header section of the sub-pattern dictionary storage unit is selected. Therefore, when the feature pattern and the standard pattern are matched by the matching unit, the number of standard patterns to be matched can be reduced, and the processing speed can be increased.

Further, the handwritten character recognition device according to claim 4 is
2. The handwritten character recognition device according to claim 1, wherein the feature extraction unit has a direction of a straight line to which each coordinate point in the coordinate data string of the normalized substroke or the substroke chain is connected in an input order is a predetermined value or more. Bending portion feature extracting means for extracting a changing bending portion as a feature point, and straight line portion feature extracting means for extracting feature points that approximate the feature points extracted by the bending portion feature extracting means at substantially equal intervals, When an off-stroke exists between the sub-strokes of the normalized sub-stroke chain, the off-stroke interpolating means interpolates the off-stroke at a feature point.

According to the handwritten character recognition device of the fourth aspect, each of the coordinate points in the coordinate data string of the normalized substroke or substroke chain is input in the input order by the bent portion feature extraction means of the feature extraction unit. A bent portion in which the direction of the connected straight line changes by a predetermined value or more is extracted as a feature point. Then, the straight line feature extraction means extracts feature points that approximate the feature points extracted by the bent part feature extraction means at substantially equal intervals. When there is an off-stroke between the sub-strokes of the normalized sub-stroke chain, the off-stroke interpolating means interpolates the off-stroke at a feature point. Therefore, since the off-stroke is interpolated when the feature is extracted, the “continuation” in the sub-pattern can be easily dealt with.

Further, the handwritten character recognition device according to claim 5 is:
The handwritten character recognition device according to claim 4, wherein the sub-pattern dictionary storage unit stores, for each of the registered standard patterns,
While having a header section in which information regarding complexity based on the number of feature points of the bent portion extracted by the bent portion feature extraction means is registered, the dictionary selection section is configured to extract the bent portion feature of the feature extracting section. Obtaining the approximate straight line number of the stroke from the number of characteristic points of the bent portion extracted by means, having a complexity limit means for defining a limited range of the complexity of the stroke based on the approximate straight line number, The dictionary selection unit selects from the sub-pattern dictionary storage unit only a standard pattern having information about complexity within the limited range defined by the complexity limiting unit in the header section of the sub-pattern dictionary storage unit. It has a feature.

According to the handwritten character recognition device of the fifth aspect, the number of characteristic points of the bent portion extracted by the bent portion feature extracting means is set for each standard pattern registered in the sub-pattern dictionary storage unit. Information on the complexity based on this is registered in the header part in advance. Then, by the complexity limiting means of the dictionary selection unit, the approximate straight line number of the stroke is obtained from the number of the characteristic points of the bent portion extracted by the bent portion feature extracting unit, and based on the approximate straight line number, Define a limited range of stroke complexity. After that, the dictionary selection unit selects from the sub-pattern dictionary storage unit only the standard pattern having in the header section information regarding the complexity within the limited range defined by the complexity limiting unit. Therefore, when matching the characteristic pattern and the standard pattern by the matching unit,
You can reduce the number of standard patterns to be matched,
The processing speed can be increased.

Further, the handwritten character recognition device of claim 6 is
The handwritten character recognition device according to claim 1, wherein the stroke segmentation unit, when dividing the stroke into the substrokes, information indicating whether a start point and an end point of the substroke are end points of the stroke, Alternatively, while having a sub-stroke cutout information generating unit that generates information indicating a condition where the start point and the end point are segmentation positions as cutout information of the substroke, the feature extraction unit,
Based on the sub-stroke cut-out information generated by the sub-stroke cut-out information generating means, the cut-out information of the start point of the first sub-stroke and the end point of the last sub-stroke forming the stroke is cut-out information of the stroke. The sub-pattern dictionary storage unit has, for each of the registered standard patterns, the same process as that performed by the sub-stroke cut-out information generation unit and the stroke cut-out information generation unit. The dictionary selection section has a header section in which the cut-out information of the cut-out standard pattern is registered, and the dictionary selection section includes the stroke cut-out information generated by the stroke cut-out information generation means and the header section of the sub-pattern dictionary storage section. Above standard pattern And clipping information is only the standard pattern that satisfies a predetermined condition is characterized by electing from the sub-pattern dictionary storing unit.

According to the handwritten character recognition device of the sixth aspect, the substroke cutout information generating means of the stroke segmentation unit determines the start point and the end point of the substroke when dividing the stroke into the substrokes. Information indicating whether or not it is the end point of the stroke, or information indicating a condition in which the start point and the end point thereof are segmentation positions is generated as the cutout information of the substroke. Next, by the stroke cutout information generation means of the feature extraction unit,
Based on the cutout information of the substroke, cutout information of the start point of the first substroke and the end point of the last substroke forming the stroke is generated as the cutout information of the stroke. On the other hand, for each standard pattern registered in the sub-pattern dictionary storage unit, the cut-out information of the standard pattern cut out in the same manner as the cut-out by the sub-stroke cut-out information generation means and the stroke cut-out information generation means is added to the sub-pattern. It is registered in advance in the header section of the pattern dictionary storage section. A standard pattern in which the stroke selection information generated by the stroke cutout information generating unit and the standard pattern cutout information in the header portion of the sub-pattern dictionary storage unit satisfy a predetermined condition by the dictionary selection unit. Elect only. Therefore, when the feature pattern and the standard pattern are matched by the matching unit, the number of standard patterns to be matched can be reduced, and the processing speed can be increased.

The program storage medium of claim 7 is
Or coordinates of the handwriting input by handwriting is the one sub-strokes strokes represented by the coordinate data strings which are arranged in time series, or dividing into two or more sub-stroke, upper Symbol sub stroke or sub A step of normalizing each sub-stroke chain having a plurality of strokes, and a characteristic relating to the normalized sub-stroke
Amount or the feature point sequence having the feature amount related to the substroke chain is the stroke of the handwriting input by handwriting.
The above-mentioned sub stroke or the above sub straw obtained from
Extracting a characteristic pattern of click chain, previously
A given character or a putter that forms part of a character
For each sub-pattern representing the coordinate data sequence of the handwriting, a feature point sequence having a feature amount related to a sub-stroke forming the sub-pattern or a feature amount related to a sub-stroke chain in which a plurality of sub-strokes are connected is handwritten as the handwriting input. of
To match with the feature pattern obtained from the stroke
Registering in the sub-pattern dictionary as a standard pattern, selecting a standard pattern satisfying a predetermined condition from the standard patterns registered in the sub-pattern dictionary, the extracted feature pattern and the sub-pattern By matching with the standard pattern selected from the pattern dictionary, the pattern similar to the characteristic pattern can be obtained.
And a step of recognizing a sub pattern candidate consisting of the standard pattern and a step of generating a handwritten input word by combining the recognized sub pattern candidates with reference to a word dictionary. It is characterized by

According to the program storage medium of the seventh aspect, a coordinate data string in which the coordinates of handwriting input by handwriting are arranged in time series is input from the input device such as the tablet, and each time one stroke ends. First, the stroke represented by the coordinate data string is divided into one or more substrokes. Then, the coordinate data string of the sub-stroke is normalized for each sub-stroke that can be a sub-pattern that forms a character or a part of a character, or for each sub-stroke chain in which a plurality of sub-strokes are connected. The sub-pattern may form a character string regardless of one character. A feature point sequence having a feature amount related to the normalized substroke or substroke chain is extracted as a feature pattern. On the other hand, a standard pattern extracted as a feature point sequence having a feature amount related to a substroke or a substroke chain of a plurality of substrokes forming a predetermined character or a part of a character is registered in advance in a subpattern dictionary. Keep it. In this way, a standard pattern satisfying a predetermined condition is selected from the standard patterns registered in the sub-pattern dictionary. Then, the extracted characteristic pattern is collated with the standard pattern selected from the sub-pattern dictionary to recognize the sub-pattern candidate. In this way, substroke division, normalization, feature extraction, standard pattern selection, and subpattern candidate recognition are sequentially repeated for each stroke to obtain a subpattern candidate forming a character or a part of a character. Then, the handwritten input word is generated by combining the sub-pattern candidates for each stroke while referring to the word dictionary.

In this way, the stroke of the handwriting input by handwriting is made one substroke or divided into two or more substrokes, so that the characters existing in the Cursive style and the mixed style strokes and the character continuation can be continued. Most of the parts are separated into individual characters. This character separation is in the Cursive style and the M
Since it is difficult to separate all the characters in the ixed style stroke, by introducing the concept of a sub-pattern that forms a character or a part of a character to cope with difficult-to-separate character chains, delayed strokes, etc.
Can recognize rsive-style and Mixed-style handwriting.

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION The handwritten character recognition apparatus of the present invention will be described in detail below with reference to the embodiments shown in the drawings. FIG. 2 is a block diagram of a handwritten character recognition device according to an embodiment of the present invention. This handwritten character recognition device is a tablet 1
, Segmentation unit 2, and feature extraction storage unit 3
The sub pattern recognition unit 4, the sub pattern dictionary storage unit 5, the word generation unit 6, the word generation storage unit 7, the word dictionary storage unit 8, the result display unit 9, and the control unit 10. .

The tablet 1 is formed, for example, of a display-integrated type, and converts handwriting data pen-input by a writer into a coordinate data string represented by an electric signal.

Further, the segmentation unit 2 has an input unit 2a to which the coordinate data sequence from the tablet 1 is input, a word segmentation unit 2b for segmenting (dividing) words, and one stroke as one substroke. Or a stroke segmentation unit 2c that divides the stroke into two or more substrokes. The stroke segmentation unit 2c has a sub-stroke cutout information generating means 17 for generating substroke cutout information.

The feature extracting storage unit 3 also includes an input coordinate buffer 3a used when extracting features of an input coordinate data string, a pointer to the input coordinate buffer 3a for each substroke, and various features. And a characteristic buffer 3b for storing characteristic patterns and a characteristic pattern buffer 3c for storing characteristic patterns.

Further, the sub-pattern recognition unit 4 normalizes a coordinate data string of one sub-stroke or a sub-stroke chain in which a plurality of sub-strokes are continuous, and a characteristic of the normalized sub-stroke chain. A feature extraction unit 4b that extracts points as a feature pattern, a dictionary selection unit 4c that selects a standard pattern registered in a sub-pattern dictionary storage unit 5 described later, and a feature pattern buffer 3c.
A matching unit 4d that matches the characteristic pattern stored in the dictionary with the standard pattern selected by the dictionary selecting unit 4c;
The sub-pattern candidate sorting unit 4e rearranges the sub-pattern candidates stored in the sub-pattern candidate buffer 7a described later. The normalization unit 4a is
The feature extraction unit 4b has a normalization condition selection unit 11 for selecting a normalization condition, and the feature extraction unit 4b uses a bent portion of a coordinate data string of a substroke or a substroke chain normalized by the normalization unit 4a as a feature point. Bending part feature extracting means 12 for extracting, straight line part extracting means 13 for interpolating the feature points extracted by the bending part feature extracting means 12 at substantially equal intervals, and off-stroke between substrokes of the substroke chain. Has an off-stroke interpolation means 14 for interpolating with the feature points, and a stroke cutout information generation means 15 for generating stroke cutout information. Further, the dictionary selection unit 4c has a complexity limiter 16 for defining a limit range of the complexity of strokes.

Further, the word generation unit 6 refers to the word dictionary registered in the word dictionary storage unit 8 and combines the sub-pattern candidates stored in the sub-pattern candidate buffer 7a to generate a word candidate. A section 6a, a word candidate sorting section 6b for rearranging word candidates stored in a word candidate buffer 7c described later, and an output section 6c for displaying the word candidates stored in the word candidate sorting section 6b in descending order of similarity. have.

Further, the word generation storage section 7 includes a sub pattern candidate buffer 7a used when a word is generated by combining sub pattern candidates, and a sub pattern candidate index for storing index information of the sub pattern candidate buffer 7a. 7b, a word candidate buffer 7c for storing the word candidates generated by the word matching unit 6a of the word generation unit 6, and a word candidate index 7d for storing the index information of the word candidate buffer 7c.

The word dictionary storage unit 8 stores a tree-structured dictionary of recognition target words, and the control unit 10 is stored in a ROM (read only memory) not shown. According to the program, the segmentation unit 2, the sub-pattern recognition unit 4, and the word generation unit 6 are controlled to execute the handwritten character recognition process.

In the handwritten character recognition device having the above structure,
The coordinate data string input from the tablet 1 is stored in the input coordinate buffer 3a via the input unit 2a. At this time, the input unit 2a detects a pen-off based on this coordinate data string and adds a stroke end code.
Then, the word segmentation unit 2b performs word segmentation using information such as the space between strokes and the time from the pen-off of the last stroke input, and adds a word end code. The stroke segmentation unit 2c detects a segmentation position estimated to be a boundary between characters from a coordinate data string of one stroke, and makes one stroke one substroke or divides it into two or more substrokes. Then, the pointer to the input coordinate buffer 3a representing the section of the coordinate data string of the one or more substrokes and various features are stored in the feature buffer 3b for each substroke.

Next, the normalizing section 4a uses the pointer to the input coordinate buffer representing the section of the coordinate data string of the substroke stored in the feature buffer 3b to form a subpattern forming a character or a part of a character. A coordinate data string of one substroke or one or more substroke chains up to the maximum number of substrokes is generated, and the substroke or substroke chain is normalized.

Next, the bent portion feature extraction means 12 of the feature extraction unit 4b extracts the coordinate points of the bent portion as feature points for the normalized substroke or substroke chain. Then, the straight line feature extraction means 13 of the feature extraction unit 4b adds / deletes feature points in order to align the distances and directions between the respective feature points extracted by the bent portion feature extraction means 12. Further, feature points are inserted assuming that there is a stroke in a section (hereinafter referred to as an off stroke) not actually being pen-inputted by the off-stroke interpolation means 14 of the feature extracting section 4b, and finally obtained. The sub-pattern (feature pattern) of the feature point sequence is stored in the feature pattern buffer 3c. The dictionary selection unit 4c is
The header information of the standard pattern registered in the sub-pattern dictionary storage unit 5 is read and the standard pattern satisfying various conditions is selected.

Next, the matching section 4d matches the characteristic pattern stored in the characteristic pattern buffer 3c with the standard pattern selected by the dictionary selecting section 4c, recognizes the sub-pattern candidates, and recognizes them. The sub pattern candidates are stored in the sub pattern candidate buffer 7a. Then, the sub pattern candidate sorting unit 4e
Thus, the index information of the sub-pattern candidate index 7b pointing to the sub-pattern candidate stored in the sub-pattern candidate buffer 7a is sorted based on the degree of similarity.

Here, the standard patterns registered in the sub-pattern dictionary storage unit 5 are the writing style and the curling pattern.
It is generated as follows so that the sive style etc. can be dealt with. In addition, in this embodiment, the definition of the sub-pattern is defined as the following (A) to (C). (A) A pattern that constitutes one character is a sub-pattern. (B) For characters in which a part of the character may be a delayed stroke, the pattern forming each part is a sub-pattern. (C) When dividing a stroke into substrokes,
A pattern that forms a portion longer than one character where the boundary between characters is difficult to divide is a sub-pattern. A sub-pattern is set according to the definition of the sub-pattern, and a stroke order difference and a deformation variation of each sub-pattern are prepared (hereinafter, the individual stroke order patterns and deformation variation patterns are also referred to as sub-patterns). Then, a feature point sequence obtained by performing the same process as the feature extraction process performed by the normalization unit 4a and the feature extraction unit 4b on the coordinate data sequence of each sub-pattern is registered in the sub-pattern dictionary storage unit 5 in advance as a standard pattern. Keep it.

Next, the word collation unit 6a refers to the word dictionary registered in the word dictionary storage unit 8 and combines the sub pattern candidates stored in the sub pattern candidate buffer 7a to generate a word candidate. To do. Then, the word candidate sorting unit 6b sorts the index information of the word candidate index 7d pointing to the word candidate stored in the word candidate buffer 7c generated by the word matching unit 6a based on the similarity.

Then, the output unit 6c sends the word candidates stored in the word candidate buffer 7c to the result display unit 9 in descending order of similarity, and displays them on the result display unit 9.

FIG. 3 is a flow chart showing a general flow of the handwritten character recognition processing of the control unit 10 of the handwritten character recognition apparatus.

The outline of the handwritten character recognition process will be described below with reference to FIG.

First, when the handwritten character recognition process starts, in step S11, the coordinate data string from the tablet 1 is fetched by the input unit 2a and stored in the input coordinate buffer 3a.

Next, in step S12, the word segmentation unit 2b determines the end of a word based on the space between strokes, the time from the pen-off of the last stroke input, and the like. Then, if it is determined in step S12 that the word has not ended, the process proceeds to step S13.

Then, in step S13, each stroke input by the stroke segmentation unit 2c is set as one sub-stroke, or 2
The sub-stroke is divided into the above sub-strokes, and a pointer to the input coordinate buffer 3a representing the section of the coordinate data string of the sub-stroke and various characteristics are stored in the characteristic buffer 3b.

Next, in step S14, the normalization section 4a, the feature extraction section 4b, the dictionary selection section 4c, and the matching section 4d are processed.
And the sub-pattern candidate sorting unit 4e,
Sub-strokes or sub-stroke chains are sequentially generated for normalization, feature extraction, sub-pattern matching and sub-pattern candidate sorting.

If it is determined in step S12 that the word ends, the process proceeds to step S15.
The word matching unit 6a combines the sub-pattern candidates stored in the sub-pattern candidate buffer 7a with reference to the word information in the word dictionary storage unit 8 to generate word candidates.

Next, in step S16, the word candidates generated by the word matching unit 6a are sorted by the word candidate sorting 6b in descending order of similarity,
The results are displayed on the result display unit 9 in descending order of similarity.

Next, the segmentation processing, sub-pattern recognition processing and word generation processing in the general flow shown in FIG. 3 will be described in detail.

[I] Segmentation Processing FIG. 4 is a flow chart of stroke segmentation processing by the stroke segmentation unit 2c.

The stroke segmentation processing will be described in detail below with reference to FIG. 4 by taking the processing of dividing a stroke into substrokes as an example.

First, when the segmentation process starts, in step S21, the direction between the coordinate data strings forming the stroke is quantized by the 16-direction code shown in FIG. 5 (a).

Next, in step S22, the direction difference HD between the direction values is calculated as follows: HD = Hdiff (H1, H2) = min (| H1-H2 |, 16-
Calculate using | H1-H2 |). However, H1 and H2 are direction values in two directions (1 ≦ H1 ≦ 16, 1 ≦ H2 ≦ 16) at a certain feature point, and min (a, b) takes the smaller value of a and b. Means that. Further, FIG. 5B is a diagram showing the relationship among the coordinate data sequence, the direction code sequence, and the direction difference sequence that form the stroke. Although the above direction values are set to 1 to 16,
The configuration is not limited to this, and may be set appropriately as needed.

Next, in step S23, the trajectory of the stroke is traced using the coordinate data sequence, the direction code sequence and the direction difference sequence forming the stroke, and a flag is set at the segmentation position satisfying the segmentation condition for cutting out as a substroke. Set. This segmentation condition is, for example, when the writing is generally made in the Cursive style, a rule of the movement of the pen in the part where the following characters are connected to each other. -When connecting to the next character from'a ',' c ',' d ', etc., the pen moves like a'U' character. -When connecting to the next character from'g ',' j ',' y ', etc., the pen moves so as to draw a self-loop.・ When writing'p ',' s ', etc., the pen moves so as to fold back ・ When connecting from the crossbar of't' to the next character,
The pen intersects the previous stroke.

Next, in step S24, the segmentation position is verified from the length of the estimated sub-stroke between the segmentation positions, the height of the coordinate data string input up to that point, and the unnecessary segmentation position is detected. Clear the flag that represents.

Next, in step S25, the characteristics of the substroke are extracted. That is, a pointer to an input coordinate buffer that represents a section of a coordinate data string of a substroke chain divided by one substroke or a segmentation position, various characteristics of the substroke (for example, maximum / minimum coordinate value, substroke length, (Cutout information, etc.) is extracted and stored in the feature buffer 3b. At this time, the sub-stroke cutout information generating means 17 of the stroke segmentation unit 2c determines whether the start point and the end point of each substroke are the end points of the stroke, or the start point and the end point thereof. Information indicating the condition as the segmentation position is generated as the above-mentioned substroke cutout information, and the substroke cutout information is stored in the feature buffer 3b. Then, this segmentation process is completed.

FIG. 16A shows an example in which the above stroke segmentation processing is performed. In FIG. 16A, the segmentation position of the handwritten character string “She told them” is indicated by black dots.

[II] Sub-Pattern Recognition Processing Next, sub-pattern recognition processing will be described. This sub-pattern recognition processing is performed on the feature pattern D extracted from the coordinate data sequence of the sub-stroke or sub-stroke chain obtained by the segmentation processing.
This is a process in which a matching distance dist is obtained by P (Dynamic Programming) matching processing, and sub-pattern candidates up to the Nth place are created in the order of decreasing matching distance dist.

FIG. 6 shows the normalization section 4a, the feature extraction section 4b, the dictionary selection section 4c, and the matching section 4 of the sub-pattern recognition section 4.
9 is a flowchart of a sub-pattern recognition process by the d and sub-pattern candidate sorting unit 4e. here,
To facilitate the explanation of the sub-pattern recognition process,
Regarding the normalization process by the normalization unit 4a, the feature extraction process by the feature extraction unit 4b, the dictionary selection process by the dictionary selection unit 4c, the DP matching process by the matching unit 4d, and the structure of the sub pattern dictionary of the sub pattern dictionary storage unit 5 in advance, I will explain.

(1) Normalization Process In the normalization process, a coordinate data string of one substroke or a coordinate data string of a substroke chain consisting of one or more substrokes up to the maximum number of substrokes is used. , The aspect ratio of the stroke is obtained from the maximum / minimum value of the coordinate values in the coordinate data sequence, and the normalized number of dots in the vertical / horizontal direction (Xnorm, Yn
orm) is selected. Then, the circumscribing frame of the coordinate data string is normalized so as to have horizontal Xnorm dots and vertical Ynorm dots. In this embodiment, the normalization condition selection means 11 of the normalization unit 4a selects from three types of normalization conditions according to the aspect ratio. The three types of normalization conditions are Xnorm = 64 dots, Ynorm = 128 dots up to an aspect ratio of 2: 1 and Xnorm = 128 dots, Ynorm = 128 dots, an aspect ratio of 1 from 2: 1 to 1: 2. : If larger than 2, Xnorm = 256 dots and Ynorm = 128 dots.

(2) Feature Extraction Process In the feature extraction process, the bent feature extractor 13 of the feature extractor 4b scans the normalized coordinate data sequence in the order of input to change the direction above the threshold value. The bending point is determined as the bending point, and the end point of the stroke is determined as the end point. Then, when the distance between the adjacent bending points or the end points is equal to or more than a predetermined value, the interpolation point is obtained by the straight line feature extracting means 13 of the feature extracting unit 4b at a constant dot interval. Similarly, when the off-stroke interpolation means 14 of the feature extraction unit 4b includes an off-stroke section in the coordinate data string and the distance of the off-stroke section is equal to or greater than a predetermined value, interpolation points are set at fixed dot intervals. Ask for. A characteristic point obtained by combining the bending point, the end point, and the interpolation point is extracted as a characteristic pattern. Then, for each feature point, the direction value d to the next feature point is calculated, and the coordinate value (x,
y) together with the feature amount Fn. Therefore, the feature amount Fn is a variable length coordinate sequence with a direction represented by the following equation. Fn = (x ₁ , y ₁ , d ₁ ), (x ₂ , y ₂ , d ₂ ), ..., (x _n , y _n ,
d _n) Incidentally, the direction value d is, for example, a quantized value by the D direction code FIGS. 5 (a) D 1 in the same manner as (FIGS. 5 (a) in 16).

(3) Dictionary Selection Processing In the dictionary selection processing, the number of approximate straight lines (hereinafter referred to as complexity) obtained from the normalization type and the number of inflection points as the normalization conditions obtained until the characteristic pattern is created. It is a process for speeding up the process of DP (Dynamic Programming) matching. By the complexity limiting means 16 of the dictionary selection unit 4c, the limited range of the complexity p is (p
It is not less than −α) and not more than (p + α) (α is an empirically determined value). That is, the standard patterns registered in the sub-pattern dictionary storage unit 5 in advance are classified according to the normalization type and then sorted in the order of the number of complexity levels. Then, in the DP matching process, not all standard patterns are used, but only standard patterns of sub-patterns having a normalization type related to the feature pattern of interest and a complexity falling within a limited range of complexity are used. The DP matching process is performed.

(4) DP Matching Process The DP matching process is a process of calculating the distance between the characteristic pattern from the stroke coordinate data string and the standard pattern registered in the sub-pattern dictionary storage unit 5. This DP matching processing is to obtain the distance between two data having different lengths. Details of the DP matching processing are described in, for example, Speech Information Engineering (NTT Technology Transfer Co., Ltd., p150-153). There is.

In this embodiment, the distance d (i, j) between the i-th feature of the feature pattern having I features and the j-th feature of the standard pattern having J features is set as the matching path. The partial sum D (i, j) of the distances along and the matching distance dist are defined by the following expressions.

[Equation 1]

(5) Structure of Sub-Pattern Dictionary In the sub-pattern dictionary storage unit 5, a standard pattern is registered in advance for each sub-pattern. The data of this standard pattern is a variable-length coordinate sequence with a direction similar to the above-mentioned feature pattern and feature amount Fn. Normalization processing by the normalization unit 4a is performed on the coordinate data sequence of the sub-pattern prepared in advance. It is created by performing the feature extraction processing by the feature extraction unit 4b. Further, in the sub-pattern dictionary storage unit 5, various information of each sub-pattern is added to the header section as header information. As shown in FIG. 9 (a), the header part stores the normalization type according to the aspect ratio when performing normalization in the normalization part 4a in the “normalization type part”, and in the “complexity part”. , And stores the degree of complexity, which is the number of bending points extracted by the feature extraction unit 4b. The dictionary structure of the sub-pattern dictionary storage unit 5 is composed of a header section and a variable-length coordinate data string for each sub-pattern dictionary.

The normalization processing by the normalization unit 4a and the feature extraction processing by the feature extraction unit 4b described in (1) to (5) above,
Based on the dictionary selection processing by the dictionary selection unit 4c, the DP matching processing by the matching unit 4d, and the structure of the sub-pattern dictionary in the sub-pattern dictionary storage unit 5, the i-th stroke is input for the sub-pattern recognition processing according to FIG. Seg by the sub-segmentation unit 2c.
The case of being divided into _num substrokes will be described.

First, in step S31, it is determined whether or not the input i-th stroke is the first stroke. As a result, when the stroke is the first stroke, step S32
If the stroke is not the first stroke, the process proceeds to step S.
Proceed to 33.

Next, in step S32, the substroke end number sstrk_en set in the word generation storage section 7 is set.
The initial value "1" is set in d, and the initial value "1" is set in the sub-pattern candidate counter sc set in the word generation storage unit 7.

Next, in step S33, the substroke start number sstrk_start set in the word generation storage unit 7 is set.
To the contents of the sub-stroke end number sstrk_end, and the stroke division number seg_num is added to the contents of the sub-stroke end number sstrk_end.

Next, in step S34, the sub-pattern start number sp_sta set in the word generation storage section 7 is set.
The value of (substroke start number sstrk_start-maximum substroke number SSTRK_MAX + 1 for performing stroke chain) or "1" is set to rt, whichever is larger.
In addition, the sub-stroke start number sstrk_st is added to the sub-pattern end number sp_end set in the word generation storage unit 7.
Set the content of art.

Next, in step S35, the content of the sub-pattern candidate counter sc is stored in the "start buffer number" of the word candidate index 7d. In the “start buffer number” of the word candidate index 7d, the “sub-pattern start number” in the word candidate index 7b is the content of the sub-pattern start number sp_start, and the “sub-stroke number” is
Is a value consisting of (contents of sub-pattern end number sp_end-sub-pattern start number sp_start + 1). The sub-pattern candidate index 7b has a two-dimensional array structure as shown in FIG. The sub-pattern candidate index 7b is a buffer that stores information about where in the sub-pattern candidate buffer 7a sub-pattern candidates consisting of a "start stroke number" to a "sub-stroke number" of sub-stroke chains are stored. is there.

Next, in step S36, N sub-pattern candidate buffers from the'sc'th of the sub-pattern candidate buffer 7a are initialized. The sub-pattern candidate buffer 7a has a structure as shown in FIG. 9 (b) for each candidate. Further, the maximum number of sub-pattern candidates per section obtained in this embodiment is N. Here, in the initialization of the sub-pattern candidate buffer 7a, the maximum value maxdist is set to all N distance parts of the sub-pattern candidate buffer for one section.

Next, in step S37, the normalization unit 4a performs the above-described normalization processing on the coordinate data string from the'sp_start'th substroke to the'sp_end'th substroke. . At this time,
The normalization type selected from the aspect ratio of the stroke composed of the coordinate data string is stored in the feature pattern buffer 3c.

Next, in step S38, the feature extraction unit 4b performs the above-described feature extraction process on the coordinate data sequence after the normalization process in step S37 to obtain a feature pattern. The characteristic pattern thus obtained is stored in the characteristic pattern buffer 3c.
At this time, the bent portion feature extraction means 1 of the feature extraction unit 4b
2, the complexity of the number of bending points is also stored in the feature pattern buffer 3c at the same time. In addition, the feature extraction unit 4b
Based on the sub-stroke cut-out information stored in the feature buffer 3b, the stroke cut-out information generation means 15 of the above-mentioned stroke cut-out information generation means 15 outputs the cut-out information of the start point of the first sub-stroke and the end point of the last sub-stroke constituting the stroke. It is generated as cutout information of the stroke.

Next, in step S39, the dictionary selecting section 4c and the matching section 4d execute the sub-pattern recognition processing subroutine described later, using the characteristic pattern buffer 3c obtained in the above steps as input data.

Next, the process proceeds to step S40, and the sub-pattern candidate sorting section 4e performs step S3.
As a result of the sub-pattern recognition processing in 7, the sub-pattern candidates stored in the sub-pattern candidate buffer 7a are sorted in ascending order of the matching distance dist.

Next, in step S41, (content of sub pattern candidate counter sc-1) is stored in the "end buffer number" of the word candidate index 7d.

Next, in step S42, the content of the sub-pattern start number sp_start is incremented.

Next, in step S43, the content of the sub pattern start number sp_start is the sub pattern end number.
Determine if it is greater than the content of sp_end. As a result, if the content of the sub-pattern start number sp_start is less than or equal to the content of the sub-pattern end number sp_end, step S35.
Then, the process goes to the sub pattern recognition process for the sub stroke chain from the sub pattern start number sp_start to the'sub pattern end number sp_end. Thus, steps S30 to S4 are performed from the contents of the sub-pattern start number sp_start to the contents of the sub-pattern end number sp_end.
Repeat 3. On the other hand, if the content of the sub-pattern start number sp_start is larger than the content of the sub-pattern end number sp_end in step S43, the process proceeds to step S44.

Next, in step S44, the content of the sub-stroke start number sstrk_start is incremented.

Next, in step S45, it is determined whether the content of the substroke start number sstrk_start is larger than the content of the substroke end number sstrk_end. As a result, if the content of the substroke start number sstrk_start is less than or equal to the content of the substroke end number sstrk_end, the process returns to step S34, the subpattern start number sp_start and the subpattern end number sp_end are set, and the subpattern start number sp_start is started. The process moves to the sub-pattern recognition process for the sub-stroke chain up to the sub-pattern end number sp_end. Thus, steps S34 to S45 are repeated from the content of the sub-stroke start number sstrk_start to the sub-stroke end number sstrk_end. On the other hand, in step S45, the content of the substroke start number sstrk_start is changed to the substroke end number sstrk_e.
If it is larger than the content of nd, this sub-pattern recognition processing is terminated.

7 and 8 are flowcharts of the sub pattern recognition processing subroutine.

The sub-pattern recognition process executed by the dictionary selection unit 4c and the matching unit 4d will be described in detail below with reference to FIGS.

First, in FIG. 7, in step S51,
The dictionary selection unit 4c causes the feature pattern buffer 3c
A start pointer dicp and an end pointer when the normalization type and the complexity stored in the sub-pattern dictionary storage unit 5 are read as follows.
Set endp and. Here, it is assumed that the sub-pattern dictionary storage unit 5 is sorted according to the normalization type first and then sorted in the order of the number of complexity levels, as described above. Then, in the sub-pattern dictionary storage unit 5, a table in which the start address of a standard pattern having the same normalization type and the same complexity registered and the normalization type and the complexity are associated with each other. And are registered as additional information. Therefore, the read start address for reading the standard pattern is obtained by referring to the table based on the normalization type and the lower limit (p−α) of the limited range of complexity. Further, based on the normalization type and the upper limit ((p-α) +1) of the limited range of complexity, the table is referred to obtain the read end address when the standard pattern is read. The read start address and end address thus obtained are set in the start pointer dicp and end pointer endp set in the word generation storage unit 7, respectively.

Next, in step S52, the maximum distance is searched from the'distance part 'in the N'th sub-pattern candidate buffer 7a from the'sc'th, and the searched maximum distance is stored in the word generation storage unit 7. Store it in the set maximum distance maxdist.

Next, in step S53, the header part is read from the address designated by the content of the start pointer dicp when reading the standard pattern from the sub-pattern dictionary storage part 5.

Next, in step S54, it is determined whether or not to skip to the next standard pattern. That is, the stroke cutout information generating means 1 of the feature extracting unit 4b.
5, based on the sub-stroke cutout information stored in the feature buffer 3b, the cutout information of the first substroke start point and the substroke end point forming the stroke is generated as the stroke cutout information. . Then, the dictionary selection unit 4d skips because the stroke cutout information and the standard pattern cutout information registered in the header portion of the sub-pattern dictionary storage unit 5 do not satisfy a predetermined condition, or , And it is determined whether or not to skip. As a result, in the case of skipping, the standard pattern is not targeted for the DP matching processing of the matching unit 4d, and the process proceeds to step S61 shown in FIG. 8 to proceed to the next standard pattern.
It proceeds to step S55. In the header part of the sub-pattern dictionary storage unit 5, the cut-out information of the standard pattern cut out in the same manner as the cut-out by the stroke cut-out information generating means 15 is registered in advance for each registered standard pattern. ing.

Then, in step S55, the matching section 4d executes the DP matching processing of the characteristic pattern and the standard pattern, and the obtained matching distance is set to the matching distance set in the word generation storage section 7. di
Store in st.

Then, the process proceeds to step S56, and step S
The content of the matching distance dist obtained in 55 is the maximum distance ma
It is determined whether it is less than or equal to the contents of xdist. as a result,
If the content of the matching distance dist is less than or equal to the content of the maximum distance maxdist, the process proceeds to step S57 shown in FIG. 8. On the other hand, if the content of the matching distance dist is larger than the content of the maximum distance maxdist, the process proceeds to step S61 shown in FIG. Go to the next standard pattern.

Next, in step S57 of FIG. 8, in the N sub-pattern buffers from the'sc'th sub-pattern candidate buffer 7a, the same sub-pattern code (character It is determined whether there is a sub pattern candidate having a code). As a result, when there is no sub-pattern candidate having the same sub-pattern code, the process proceeds to step S58,
If there are sub-pattern candidates having the same sub-pattern code, the process proceeds to step S59.

Next, in step S58, the maximum distance ma
After rewriting the information of the sub pattern candidate buffer 7a presenting xdist to the information of the standard pattern, the process proceeds to step S60. This information is rewritten as follows. That is, the'distance part 'stores the contents of the matching distance dist, and the'code part' stores the sub-pattern code of the header part.

On the other hand, in step S59, the sub-pattern candidate buffer 7a in which the content of the matching distance dist exhibits the same sub-pattern code as the header part of the standard pattern.
If it is smaller than the “distance part” of, the information of the sub pattern candidate buffer 7a is rewritten to the information of the standard pattern as described above, and then the process proceeds to step S60.

Next, in step S60, the maximum distance is searched from the N'distance parts' from the'sc'th position of the sub-pattern candidate buffer 7a, and the searched maximum distance is set in the word generation storage unit 7. Stored in the maximum distance maxdist.

Next, in step S61, the start pointer dicp
Increments the contents of

Next, in step S62, it is determined whether the content of the start pointer dicp is greater than or equal to the content of the end pointer endp. As a result, if it is determined that the content of the start pointer dicp is smaller than the content of the end pointer endp, step S
Returning to 53, the process shifts to the DP matching process using the next standard pattern for the characteristic pattern.

On the other hand, in step S62, the start pointer dicp
If it is determined that the content of is more than the content of the end pointer endp, the process proceeds to step S63 and the sub-pattern candidate counter
Add the number of recognized subpattern candidates to the contents of sc.
Then, this sub-pattern recognition process is terminated, and the process returns to the main routine of FIG.

When the sub pattern recognition process is completed in this way, a maximum of N sub pattern candidates consisting of standard patterns similar to the characteristic pattern are stored in the sub pattern candidate buffer 7a.

[III] Word Generation Processing Finally, the word generation processing will be described. The word generation process is a process of generating words by sequentially combining the sub-pattern candidates obtained by the sub-pattern recognition process. FIG. 11 is a flowchart of word generation processing by the word matching unit 6 and the word candidate sorting unit 6b.

The word generation process will be described in detail below with reference to FIG.

First, the word segmentation section 2c.
When the end of the word is determined by, and the stroke segmentation process and the sub-pattern recognition process are completed for all the input strokes, the word generation process starts, and in step S71, the word candidate index 7d and the word candidate buffer 7c. Are associated with each other. In this association, the buffer numbers are sequentially set from 1 to Nw in order to associate the word candidate buffer 7c with the "buffer number part" of the word candidate index 7d in a one-to-one correspondence. Here, the word candidate buffer 7c and the word candidate index 7d have a structure in which Nw pieces of the configurations shown in FIGS. 15A and 15B are arranged side by side.

Next, in step S72, a word candidate buffer to be described later is selected according to the subpattern candidate whose "start number part" is "1" among the subpattern candidates stored in the subpattern candidate buffer 7a. Execute a subroutine to initialize 7c.

Next, in step S73, the word candidate index 7d from 1 to the contents of the current word candidate counter wc set in the word generation storage unit 7 is associated with the word candidate index 7d'end number part. The minimum end number is searched for from among ', and the searched minimum end number is stored in the minimum end number min_eline set in the word generation storage unit 7.

Next, in step S74, it is determined whether or not the content of the minimum end number min_eline is smaller than the number of input substrokes (the total number of substrokes obtained by dividing each stroke). As a result, the minimum end number min_eline
If the content of is smaller than the number of input substrokes, the process proceeds to step S75, while if the content of the minimum end number min_eline is greater than or equal to the number of input substrokes, the process proceeds to step S79.

Next, in step S75, the word candidate buffer 7c is searched again, the word candidate buffer 7c having the minimum end number is selected, and the word expansion subroutine described in detail later is executed.

Next, in step S76, the word candidate index 7d is compressed. That is, from the 1st to'wc '
The second word candidate index 7d is sorted in ascending order of the contents of the'distance part ', and then the value of the'distance part' of the word candidate index 7d is the value of maxdist, and the contents of the current word candidate counter wc are counted. Subtract.

Next, in step S77, it is determined whether the content of the current word candidate counter wc is less than or equal to the maximum word expansion number WC_MAX. As a result, if the content of the current word candidate counter wc is less than or equal to the maximum word expansion number WC_MAX, the process returns to step S74 and the content of the current word candidate counter wc is the maximum word expansion number WC_MAX.
If it is larger than _MAX, the process proceeds to step S78, and the current word candidate counter wc is set to the maximum word expansion number WC_MAX. After that, the process returns to step S73, and the word expansion process is repeated until the minimum end number becomes equal to the input substroke.

On the other hand, in step S79, a final word check is performed. That is, the contents of the word dictionary indicated by the "dictionary pointer part" of the word candidate buffer 7c associated with the word candidate indexes 7d from the 1st to the "wc" th are checked, and the word candidates whose end of the word is not confirmed are deleted. Of.

Next, in step S80, word candidates are sorted. That is, the word candidate index 7d associated with the word candidate buffer 7c whose end of the word has been confirmed can be sorted in ascending order of the contents of the "distance part". After that, this word generation process is terminated.

FIG. 12 is a flow chart of a subroutine for initializing the word candidate buffer 7c in step 72 of FIG.

The word candidate buffer initialization operation will be described in detail below with reference to FIG.

First, in step S81, an initial value "1" is set in the current word candidate counter wc set in the word generation storage unit 7.

Next, in step S82, the substroke number counter c_ set in the word generation storage section 7 is set.
Set the initial value "1" to sstrk.

Next, at step S83, the content of the substroke number counter c_sstrk indicates that the maximum number of substrokes SSTRK_MAX for performing a substroke chain with a subpattern.
It is determined whether or not the following. As a result, the content of the substroke number counter c_sstrk is the maximum substroke number SSTRK_MAX.
In the following cases, while the process proceeds to step S84, the content of the sub-stroke number counter c_sstrk is the maximum sub-stroke number SST.
If it is larger than RK_MAX, the subroutine for initializing the word candidate buffer 7c is ended, and the process returns to the main routine of FIG.

Next, in the start buffer number spb_start and end buffer number spb_end set in the word generation storage unit 7 in step S84, the sub pattern start number in the sub pattern candidate index 7b is "1", and the number of sub strokes. Is the contents of the sub-stroke number counter c_sstrk, the contents of the'start buffer number 'and'end buffer number' of the sub-pattern candidate index 7b are stored.

Next, in step S85, the contents of the start buffer number spb_start is the end buffer number spb_end.
It is determined whether or not the content is less than or equal to the content. As a result, the contents of the start buffer number spb_start are changed to the end buffer number spb_end.
If the content is less than or equal to, the process proceeds to step S86.

On the other hand, in step S85, the start buffer number
If the content of spb_start is larger than the content of the end buffer number spb_end, the process proceeds to step S90, the content of the substroke number counter c_sstrk is incremented, and the process returns to step S83.

Next, in step S86, it is determined whether or not the word book is OK. That is, start buffer number spb_star
It is determined whether or not a word that starts with the sub-pattern code of the sub-pattern candidate indicated by t as a character code string exists in the word dictionary of the word dictionary storage unit 8. As a result, when it is determined that the word exists in the word dictionary of the word dictionary storage unit 8, the word dictionary is OK and the process proceeds to step S87, while when it is determined that the word does not exist in the word dictionary of the word dictionary storage unit 8, Skip to step S89.

Next, in step S87, the word candidate buffer 7c corresponding to the'wc'th word candidate index 7d is initialized. This initialization is performed as follows. That is, in FIG. 15A, in the'code part ', the sub-pattern code of the sub-pattern candidate indicated by the start buffer number spb_start is converted into a character code string,
Store from the beginning of'code part '. In'word length part ',
The number of stored characters is stored. In the'distance section ',
The value of the'distance part 'of the sub-pattern candidate is stored.
Start buffer number spb_star at the beginning of the "SP pointer part"
The value of t is stored, and “1” is stored in the “sub-pattern number part”. Further, the value of the “end number part” of the sub-pattern candidate is stored in the “end number part”. Further, the value of the "distance part" of the sub-pattern candidate is also stored in the "distance part" of the above "wc" th word candidate index 7d. And
The word dictionary storage unit 8 whose matching has been advanced to the “dictionary pointer unit”
Stores a pointer to.

Next, in step S88, the current word candidate counter wc is incremented.

Then, the process proceeds to step S89, the content of the start buffer number spb_start is incremented, and the process returns to step S85.

13 and 14 show step 75 in FIG.
3 is a flowchart of a word expansion subroutine.

The word expansion processing will be described in detail below with reference to FIGS. 13 and 14. In this word expansion process, the contents of the word candidate buffer 7c and the sub-pattern candidate buffer 7a are input, and a word dictionary having a tree structure stored in the word dictionary storage unit 8 is referenced to create a new word candidate. Processing. In step S74 of the word generation processing flowchart shown in FIG. 11, when it is determined that the minimum end number is smaller than the number of input substrokes, this word expansion subroutine is started.

First, in step S91, the new word candidate counter new_wc set in the word generation storage unit 7 is set to
The content of the current word candidate counter wc (the number of word candidates stored in the word candidate buffer 7c) +1 is set. Also,
Word candidate counter set in the word generation storage unit 7
Set the initial value '1' to wc0.

Next, in step S92, it is determined whether the content of the word candidate counter wc0 is less than or equal to the content of the current word candidate counter wc. As a result, if the content of the word candidate counter wc0 is less than or equal to the content of the current word candidate counter wc, the process proceeds to step S93. On the other hand, step S92
And the content of the word candidate counter wc0 is the current word candidate counter wc
If it is larger than the content of, go to step S108,
New word candidate counter new_ in current word candidate counter wc
After storing the contents of wc, the word expansion subroutine is terminated and the process returns to the main routine of FIG.

Next, in step S93, it is determined whether or not the word candidate has the attention end substroke number. That is, the word candidate buffer 7c indicated by the'wc0'th word candidate index 7d in the word candidate index 7d.
It is determined whether or not the contents of the'end number part '(that is, the end substroke number) is equal to the contents of the minimum end stroke number min_eline (hereinafter referred to as the noticed end substroke number). As a result, if the content of the “end number part” is equal to the content of the minimum end stroke number min_eline, the process proceeds to step S94. On the other hand, step S93
The content of'end number part 'is the minimum end stroke number min_el
If the contents of ine are not equal, the process proceeds to step S107, the maximum distance maxdist is set to the'distance part 'of the'wc0'th word candidate index 7d (invalid), and the contents of the word candidate counter wc0 are incremented. , Step S92
Return to.

Next, in step S94, the initial value "1" is substituted into the sub-stroke number counter c_sstrk.

Next, in step S95, the content of the substroke number counter c_sstrk indicates that the maximum substroke number SSTRK_ for performing substroke chaining with the subpattern.
Determine whether it is less than MAX. As a result, if the content of the sub-stroke number counter c_sstrk is less than or equal to the content of the maximum sub-stroke number SSTRK_MAX, the process proceeds to step S96, while if the content of the sub-stroke number counter c_sstrk is larger than the content of the maximum sub-stroke number SSTRK_MAX,
It proceeds to step S107.

Next, in step S96, the start buffer number spb_start and the end buffer number spb_end have the sub-pattern start number in the sub-pattern candidate index 7b as'minimum end number min_eline + 1 'and the sub-stroke number is the sub-stroke number counter c_sstrk. The contents of the'start buffer number 'and the'end buffer number' of the sub-pattern candidate index 7b are stored respectively.

Next, in step S97 shown in FIG. 14, it is determined whether the content of the start buffer number spb_start is less than or equal to the content of the end buffer number spb_end.
As a result, when the content of the start buffer number spb_start is less than or equal to the content of the end buffer number spb_end, step S98
Proceed to. On the other hand, if the content of the start buffer number spb_start is larger than the content of the end buffer number spb_end, the process proceeds to step S106, the content of the sub-stroke number counter c_sstrk is incremented at step S106, and the process returns to step S95 of FIG.

Next, in step S98, it is determined whether or not the sub pattern code of the sub pattern candidate SPB indicated by the start buffer number spb_start is a delayed stroke. As a result, if it is determined that the stroke is delayed, the process proceeds to step S100, whereas if it is determined that the stroke is not delayed, the process proceeds to step S99.

Next, in step S99, the word dictionary is OK.
Determine whether or not. That is, the sub pattern code of the sub pattern candidate SPB is converted into a character code string by referring to the word dictionary portion indicated by the “dictionary pointer section” of the word candidate WB0 corresponding to the'wc0'th word candidate index 7d. It is determined whether or not the created item can be expanded. As a result, when it is determined that the expansion is possible, the process proceeds to step S101, while when it is determined that the expansion is not possible,
Skip to step S105.

On the other hand, in step S100, it is determined whether the sub-pattern code of the sub-pattern candidate SPB indicated by the start buffer number spb_start is a pair stroke.
That is, it is determined whether or not a sub-pattern candidate that can be combined with a delayed stroke exists in the sub-pattern candidates indirectly stored in the'SP number part 'of the word candidate WB0. As a result, if the stroke is a pair stroke, the process proceeds to step S101. If the stroke is not the pair stroke, the process proceeds to step S105.

Next, in step S101, new word candidates are created in the word candidate buffer WB corresponding to the'new_wc'th word candidate index 7d. The creation of this new word candidate is performed as follows. First, each content of the word candidate WB0 is copied to the word candidate buffer WB.
The sub-pattern code of the sub-pattern candidate SPB is converted into a character code string and stored in the position indicated by the content of the “word length part” of the “code part” shown in FIG. , Add the number of stored characters. The value of the "distance part" of the sub-pattern candidate SPB is added to the "distance part". The contents of the start buffer number spb_start are stored in the position indicated by the contents of the "sub-pattern number part" of the "SP pointer part", and the contents of the "sub-pattern number part" are incremented. Further, the “end number part” stores the contents of the “end number part” of the sub-pattern candidate SPB. Furthermore, the'distance part 'of the'new_wc'th word candidate index 7d is ((distance part of word candidate buffer WB.
Content / 'sub-pattern number part' of word candidate buffer WB
Content) and store it as the average distance.

Next, in step S102, the content of the new word candidate counter new_wc is incremented.

Next, in step S103, the content of the new word candidate counter new_wc is the word candidate buffer number Nw.
It is determined whether or not the above. As a result, the new word candidate counter
If the content of new_wc is less than the number of word candidate buffers Nw, the process proceeds to step S105 while the new word candidate counter
If the content of new_wc is equal to or more than the number of word candidate buffers Nw, the process proceeds to step S104.

Next, in step S104, the word candidate buffer 7c is compressed. That is, the word candidate indexes 7d from the ('wc' + 1) th to the'new_wc'th are sorted into the contents of the'distance part 'in ascending order. Then, after ((contents of current word candidate counter wc + contents of new word candidate counter new_wc) / 2) is calculated and stored in the new word candidate counter new_wc, the process proceeds to step S105.

Next, in step S105, the content of the start buffer number spb_start is incremented, and in step S9
Return to 7.

In this way, [I] segmentation processing, [II] sub-pattern recognition processing and [III]
When word generation processing is performed and word candidates are formed in the word candidate buffer 7c, the word candidates are sent from the output unit 6c to the result display unit 9 and the word candidates are displayed on the result display unit 9.

For example, FIG. 16B shows an example in which the above sub pattern is recognized. In FIG. 16B, a sub-pattern candidate is obtained for each of five sub-strokes in which the one-stroke character string “told” is divided at the segmentation position (indicated by a black dot). The sub-pattern candidates for the first sub-stroke are l (24), t_ (26), i_ (26), e (30) in ascending order of similarity, with the notation followed by'_ '. Indicates a sub-pattern that constitutes a character portion of a character that requires a delayed stroke. Similarly, for other substrokes, o (30), v (42), l (32), e (34), t (38) d (20), a (36)-(31), _t ( 31), and the notation preceded by'_ 'indicates a sub-pattern that constitutes a delayed stroke portion of a character that requires a delayed stroke. Further, the sub-stroke chain is w (32), br (45) ol (37), d (54) in order of decreasing similarity. By combining these sub-pattern candidates with reference to the word dictionary, they are surrounded by an ellipse.
A word candidate “told” is generated by combining t_ (26), o (30), l (32), d (20), and _t (31).

As described above, the stroke of the handwriting input by handwriting is set as one substroke, or is divided into two or more substrokes to set the cursive stroke.
Most of the characters present in the style and mixed style strokes and the portion that follows the characters are separated into individual characters. Furthermore, by collating a sub-pattern that constitutes a character or a part of a character with a sub-pattern dictionary, it becomes possible to deal with a character chain or a delayed stroke that is difficult to separate. Therefore, the handwritten character recognition device described above is based on the Priting style, the Cursive style, and the Mi
Can recognize words written in xed style.

Further, a substroke or substroke chain of a subpattern forming a predetermined character or a part of the character is normalized by any one of the same normalization conditions as the normalization section 4a, and Since the standard pattern extracted as the feature point sequence having the feature amount related to the normalized substroke or the substroke chain is registered in the subpattern dictionary storage unit 5, the matching unit 4 is used.
With d, it is sufficient to match only the standard pattern of the same normalization condition as the normalization condition for the above feature pattern,
The efficiency of processing can be improved.

Further, since the bent portion feature extraction means 12, the straight line portion feature extraction means 13 and the off stroke interpolation means 14 of the feature extraction portion 4b interpolate off strokes between adjacent sub strokes at the time of feature extraction,
The "continuation" in the sub-pattern can be easily dealt with.

Further, since the dictionary selecting section 4c selects only the standard pattern having the header section with the information regarding the complexity within the limited range defined by the complexity limiting means 16 from the sub pattern dictionary storing section 5, Matching part 4
When the characteristic pattern by d is matched with the standard pattern, the number of standard patterns to be matched is reduced, and the processing can be sped up.

Further, the dictionary selection unit 4c has predetermined information on the stroke cutout information generated by the stroke cutout information generating means 15 and the standard pattern cutout information on the header portion of the sub-pattern dictionary storage unit 5. Since only the standard patterns satisfying the conditions are selected from the sub-pattern dictionary storage unit 5, the number of standard patterns to be matched can be reduced when the characteristic pattern is matched with the standard pattern by the matching unit 4d. Further, the processing speed can be further increased.

In the above embodiment, recognition of handwritten characters in English has been described, but it goes without saying that the present invention may be applied to recognition of handwritten characters in other languages.

In the above embodiment, the control unit 1
0 controls the segmentation unit 2, the sub-pattern recognition unit 4 and the word generation unit 6 by a program stored in a ROM (not shown) or the like, but a part or all of the program according to the present invention is stored in a program storage medium such as a floppy disk. It may be stored, and if necessary, part or all of the above program may be read into a device such as a personal computer and executed.

[0147]

As is apparent from the above, in the handwritten character recognition device of the invention of claim 1, a coordinate data string in which the coordinates of the handwriting input by handwriting is arranged in a time series from the input device such as the tablet. Each time one stroke is input, the stroke represented by the coordinate data string by the stroke segmentation unit is set as one substroke or divided into two or more substrokes, and the stroke segmentation unit The normalization unit normalizes the sub-stroke coordinate data string obtained by the sub-strokes, which can be a sub-pattern forming a character or a part of a character, or a sub-stroke chain in which a plurality of sub-strokes are connected, and a feature extraction unit Is the sub-stroke or sub-stroke normalized by the normalization unit The feature point sequence having a characteristic quantity of the strand extracted as feature pattern, the dictionary selection unit, and selects the reference pattern satisfying a predetermined condition from among the reference pattern registered in advance in the sub-pattern dictionary storing unit,
The matching unit compares the feature pattern extracted by the feature extraction unit with the standard pattern selected from the sub-pattern dictionary storage unit by the dictionary selection unit, and a sub-pattern candidate forming a character or a part of the character. Is recognized, and the word generation unit combines the sub-pattern candidates for each stroke while referring to the word dictionary, thereby generating a word input by handwriting.

Therefore, according to the handwritten character recognizing device of the invention of claim 1, the stroke of the handwriting input by handwriting is set as one substroke or is divided into two or more substrokes to make the Cursive style. In addition, most of the characters existing in a mixed style stroke and the part that follows the character are separated into individual characters, and in order to deal with difficult-to-separate character chains and delayed strokes, a character or part of a character is separated. Pritin by introducing the concept of subpatterns that compose
Can recognize words written in g style, Cursive style, Mixed style, etc.

Further, the handwritten character recognition device according to the invention of claim 2 is the handwritten character recognition device according to claim 1, wherein the normalizing unit is responsive to the aspect ratio of the coordinate data string of the substroke or the substroke chain. And a normalization condition selection unit that selects any one of the plurality of normalization conditions, and the normalization unit normalizes with the normalization condition selected by the normalization condition selection unit. Therefore, it is possible to prevent extreme deformation due to normalization of the input pattern.

The handwritten character recognition device of the invention of claim 3 is the handwritten character recognition device of claim 2, wherein the sub-pattern dictionary storage section creates a standard pattern for each of the registered standard patterns. The dictionary selection unit has a header part in which information on the normalization condition used at that time is registered, and the dictionary selection unit stores only the standard pattern having the information on the normalization condition in the header part of the sub-pattern dictionary storage unit. Since the pattern is selected from the pattern dictionary storage unit, when the characteristic pattern is compared with the standard pattern by the matching unit, the number of standard patterns to be compared can be reduced, and the efficiency of the process can be improved.

According to a fourth aspect of the present invention, in the handwritten character recognition device according to the first aspect, the bent portion feature extraction means of the feature extraction unit is the normalized substroke or the substroke. Extraction is made of a bent portion in which the direction of a straight line connecting each coordinate point in the coordinate data string of the stroke chain in the input order changes by a predetermined value or more, and the straight line feature extraction means of the feature extraction unit is a bent portion feature extraction. If the off-stroke exists between the sub-strokes of the normalized sub-stroke chain by extracting the feature points that approximate the feature points extracted by the means at substantially equal intervals, the off-stroke interpolation means Since the off-stroke is interpolated by the feature points, it is possible to easily deal with "continuation" in the sub-pattern.

Further, the handwritten character recognition device according to the invention of claim 5 is the handwritten character recognition device according to claim 4, wherein for each standard pattern registered in the sub-pattern dictionary storage unit,
Information on the complexity based on the number of feature points of the bent portion extracted by the bent portion feature extraction means is registered in advance in the header portion, and the bent portion feature extraction is performed by the complexity limiting means of the dictionary selection portion. Obtain the approximate straight line number of the stroke from the number of characteristic points of the bent portion extracted by means, after defining a limited range of the complexity of the stroke based on the approximate straight line number, by the dictionary selection unit, Since only the standard pattern having the information on the complexity within the limited range defined by the complexity limiting means in the header part is selected from the sub-pattern dictionary storage part, the characteristic pattern and the standard pattern by the matching part are When collating, the number of standard patterns to be collated can be reduced, and the processing speed can be increased.

Further, in the handwritten character recognition device according to the invention of claim 6, in the handwritten character recognition device according to claim 1, the substroke cutout information generating means of the stroke segmentation unit divides the stroke into the substrokes. , The information indicating whether or not the start point and the end point of the substroke are the end points of the stroke, or the information indicating the condition that the start point and the end point are the segmentation positions is generated as the cutout information of the substroke. , Based on the cutout information of the substroke, the stroke cutout information generating means of the feature extraction unit, cutout information of the start point of the first substroke and the end point of the last substroke of the stroke, the cutout information of the stroke. While generating as cutout information For each standard pattern registered in the sub-pattern dictionary storage unit, the cut-out information of the standard pattern cut out in the same manner as the cut-out by the sub-stroke cut-out information generation means and the stroke cut-out information generation means is added to the sub-pattern dictionary. Registered in advance in the header section of the storage section, the dictionary selection section stores the stroke cutout information generated by the stroke cutout information generating means and the standard pattern cutout information in the header section of the subpattern dictionary storage section. Since only standard patterns satisfying a predetermined condition are selected, the number of standard patterns to be matched can be reduced when matching the characteristic pattern with the standard pattern by the matching unit, and the processing speed can be increased. Can be planned.

According to the program storage medium of the invention of claim 7, a coordinate data string in which the coordinates of handwriting input by handwriting are arranged in time series is input from the input device such as the tablet, and the end of one stroke is completed. Each time, the stroke represented by the coordinate data string is set as one substroke, or divided into two or more substrokes,
The coordinate data string of the sub-stroke is normalized for each sub-stroke or a sub-stroke chain in which a plurality of sub-strokes can be a sub-pattern that forms a character or a part of a character, and the normalized sub-stroke or sub-stroke chain A characteristic point sequence having a characteristic amount according to the above is extracted as a characteristic pattern, and a standard pattern satisfying a predetermined condition is selected from among the standard patterns registered in the sub-pattern dictionary in advance, and the extracted characteristic pattern and sub-patterns are selected. By matching with a standard pattern selected from the pattern dictionary, recognize sub-pattern candidates that form a character or a part of a character, and for each stroke, by combining the sub-pattern candidates while referring to the word dictionary, This is to generate a word input by handwriting.

Therefore, according to the program storage medium of the invention of claim 7, the stroke of the handwriting input by handwriting is set as one substroke or is divided into two or more substrokes to achieve the Cursive style and Most of the characters existing in the mixed style stroke and the part that follows the character are separated into individual characters, and in order to deal with character chains and delayed strokes that are difficult to separate, the character or part of the character is Pritin by introducing the concept of constituent subpatterns
Can recognize words written in g style, Cursive style, Mixed style, etc.

[Brief description of drawings]

FIG. 1A to FIG. 1C are diagrams showing examples of handwritten characters of three different writing types input to a handwritten character recognition device.

FIG. 2 is a block diagram of a handwritten character recognition device according to an embodiment of the present invention.

FIG. 3 is a flowchart showing a general flow of a handwritten character recognition operation of a control unit of the handwritten character recognition device.

FIG. 4 is a flowchart of a segmentation process of the control unit.

5 (a) and 5 (b) are diagrams illustrating direction quantization between coordinate points.

FIG. 6 is a flowchart of a sub-pattern recognition process of the control unit.

FIG. 7 is a flowchart of a sub-pattern recognition subroutine of the control unit.

FIG. 8 is a flowchart of the sub-pattern recognition subroutine following the flowchart of FIG.

9A is a diagram showing a structure of a header portion of a sub-pattern dictionary storage unit of the handwritten character recognition device, and FIG. 9B is a structure of a subpattern candidate buffer of the handwritten character recognition device. FIG.

FIG. 10 is a diagram illustrating a structure and a relationship of a sub pattern candidate buffer and a sub pattern candidate index of the handwritten character recognition device.

FIG. 11 is a flowchart of a word generation process of the handwritten character recognition device.

FIG. 12 is a flowchart of a word candidate buffer initialization subroutine of the handwritten character recognition device.

FIG. 13 is a flowchart of a word expansion subroutine of the handwritten character recognition device.

FIG. 14 is a flowchart of the word expansion subroutine following the flowchart of FIG.

FIG. 15 (a) is a diagram showing a structure of a word candidate buffer of the handwritten character recognition device, and FIG. 15 (b) is a diagram showing a structure of a word candidate index.

FIG. 16 is a diagram showing an example of stroke segmentation and sub-pattern recognition of the handwritten character recognition device.

[Explanation of symbols]

1 ... Tablet, 2 ... Segmentation unit, 2a ... Input unit, 2b ... Word segmentation unit, 2c ... Stroke segmentation unit, 3 ... Feature extraction storage unit, 3a
Input coordinate buffer, 3b ... Feature buffer, 3c ... Feature pattern buffer, 4 ... Sub pattern recognition unit, 4a ... Normalization unit, 4b ... Feature extraction unit, 4c ... Dictionary selection unit, 4d ... Matching unit, 4e ... Sub pattern Candidate sorting department, 5
... sub-pattern dictionary storage section, 6 ... word generation section, 6a ... word matching section, 6b ... word candidate sorting section, 6c ... output section, 7 ... word generation storage section, 7a ... sub pattern candidate buffer, 7b ... sub pattern Candidate index, 7c ... Word candidate buffer, 7d ... Word candidate index, 8 ... Word dictionary storage section, 9 ... Result display section, 10 ... Control section, 11 ...
Normalization condition selection means, 12 ... Bent portion feature extraction means, 13
... linear part feature extraction means, 14 ... off-stroke interpolation means, 15 ... stroke cutout information generation means, 16 ... complexity degree limitation means, 17 ... substroke cutout information generation means.

Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (7)

(57) [Claims] 1. A stroke segmentation unit for dividing a stroke represented by a coordinate data string in which the coordinates of handwriting input by handwriting are arranged in time series into one substroke or dividing the stroke into two or more substrokes. When a normalizing unit that the stroke segmentation <br/> SL sub stroke or the sub-strokes on obtained by the unit normalizes each sub stroke chain more continuous, the sub stroke normalized by the normalization unit To
A feature point sequence having such a feature amount or a feature amount related to the substroke chain is stored in the stroke of the handwriting input by handwriting.
Sub stroke or sub obtained from rooke
A feature extraction unit for extracting as a feature pattern of stroke chain path which forms a part of a predetermined character or character prepared in advance
For each sub-pattern representing the coordinate data strings turn, feature points having a feature amount relating to sub-stroke chain feature amount also <br/> according to sub-stroke continuous multiple sub strokes that make up the sub <br/> Bupatan The row is the brush that was entered by handwriting above
It is possible to match with the feature pattern obtained from the trace stroke.
A sub-pattern dictionary storage unit which is registered as a standard pattern of order, and the dictionary selection unit for selecting a standard pattern that satisfies a predetermined condition from among the reference pattern registered in the sub-pattern dictionary storage unit, the feature extraction By comparing the characteristic pattern extracted by the unit with the standard pattern selected from the sub-pattern dictionary storage unit by the dictionary selection unit,
By combining a matching unit that recognizes a sub-pattern candidate composed of the standard pattern similar to the characteristic pattern and the sub-pattern candidate recognized by the matching unit with reference to a word dictionary,
A handwritten character recognition device comprising: a word generation unit that generates a word input by handwriting. 2. The handwritten character recognition device according to claim 1, wherein the normalization unit selects one of a plurality of normalization conditions according to an aspect ratio of a coordinate data string of the substroke or the substroke chain. A handwritten character recognition device having normalization condition selection means for selecting any one of them, wherein the normalization section performs normalization under the normalization condition selected by the normalization condition selection means. . 3. The handwritten character recognition device according to claim 2, wherein the sub-pattern dictionary storage section relates to the normalization condition used when creating the standard pattern for each of the registered standard patterns. The dictionary selection unit has a header part in which information is registered, and the dictionary selection unit selects only a standard pattern having information on the normalization condition in the header part of the sub-pattern dictionary storage unit from the sub-pattern dictionary storage unit. Characterized handwritten character recognition device. 4. The handwritten character recognition device according to claim 1, wherein the feature extraction unit connects the coordinate points in the coordinate data string of the normalized substroke or the substroke chain in the input order. Bending portion feature extracting means for extracting a bending portion in which the direction of a straight line changes by a predetermined value or more as a feature point, and feature points that approximate the feature points extracted by the bending portion feature extracting means at substantially equal intervals. Handwriting characterized by having straight-line feature extraction means and off-stroke interpolation means for interpolating the off-stroke at feature points when an off-stroke exists between the sub-strokes of the normalized sub-stroke chain Character recognizer. 5. The handwritten character recognition device according to claim 4, wherein the sub-pattern dictionary storage unit extracts, for each of the registered standard patterns, a feature point of the bend portion extracted by the bend feature extracting unit. While having a header section in which information on complexity based on the number of is registered, the dictionary selection section is configured to perform the stroke based on the number of characteristic points of the bent section extracted by the bent section feature extraction means of the feature extraction section. And has a complexity limiting means for determining a limited range of the complexity of the stroke based on the approximate number of straight lines, and the dictionary selecting section is provided in the header section of the sub-pattern dictionary storage section. Only the standard pattern having the information about the complexity within the limited range defined by the complexity limiting means is selected from the sub-pattern dictionary storage unit. Handwritten character recognition device. 6. The handwritten character recognition device according to claim 1, wherein when the stroke segmentation unit divides the stroke into the substrokes, a start point and an end point of the substroke are end points of the stroke. Information indicating whether or not, or with the sub-stroke cutout information generating means for generating, as the cutout information of the substroke, the information indicating the condition where the start point and the end point are segmentation positions, and the feature extraction unit, Based on the sub-stroke cut-out information generated by the sub-stroke cut-out information generating means, the cut-out information of the start point of the first sub-stroke and the end point of the last sub-stroke forming the stroke is cut-out information of the stroke. Strike as The sub-pattern dictionary storage unit, for each registered standard pattern, is cut out in the same manner as the cut-out by the sub-stroke cut-out information generation unit and the stroke cut-out information generation unit. And a header section in which cut-out information of the standard pattern is registered, and the dictionary selection section includes the stroke cut-out information generated by the stroke cut-out information generating means and the header section of the sub-pattern dictionary storage section. A handwriting character recognition device characterized in that only a standard pattern whose cut-out information of the standard pattern satisfies a predetermined condition is selected from the sub-pattern dictionary storage unit. 7. A dividing coordinates of handwriting input by handwriting is either a single sub strokes strokes represented by the coordinate data strings which are arranged in time series, or two or more sub-stroke, upper A step of normalizing each substroke or a substroke chain in which a plurality of substrokes are connected, and a feature point having the normalized feature quantity of the substroke or the feature quantity of the substroke chain The column is obtained from the strokes of the handwritten input above.
The above substroke or the above substroke chain
Extracting as a feature pattern, path constituting a part of a predetermined character or character prepared in advance
For each sub-pattern representing the coordinate data strings turn, feature points having a feature amount relating to sub-stroke chain feature amount also <br/> according to sub-stroke continuous multiple sub strokes that make up the sub <br/> Bupatan The column is the above handwritten brush
It is possible to match with the feature pattern obtained from the trace stroke.
For registering a standard pattern as a standard pattern for the sub-pattern dictionary, selecting a standard pattern satisfying a predetermined condition from the standard patterns registered in the sub-pattern dictionary, and the extracted characteristic pattern By comparing with the standard pattern selected from the sub-pattern dictionary, the standard pattern similar to the characteristic pattern
A program having a step of recognizing a subpattern candidate consisting of and a step of generating a handwritten input word by combining the recognized subpattern candidates with reference to a word dictionary. Characteristic program storage medium.